#!/usr/bin/env python3
"""Diagnose a regression model from its predictions: where it is biased, where errors grow, what it misses.
Input CSV with y_true and y_pred (names configurable), optionally a segment column, extra feature
columns to slice by, and lower/upper prediction-interval columns. Reports:
metrics MAE, RMSE, MedAE, bias (mean error), R2, WAPE
calibration mean prediction vs mean actual per decile of the prediction (a well-calibrated
regressor has ratio ~1.0 in every decile; systematic under-prediction of large
values is the most common failure)
spread correlation of |residual| with the prediction (heteroscedasticity): > 0.3 means
errors grow with the level; model log(y) or use a Gamma/Tweedie loss, report WAPE
segments MAE and bias per --segment value and per quartile of each --slice-by column
intervals coverage and mean width of [--lower, --upper] vs --nominal
worst rows the --top largest absolute errors with their row number
gates --max-abs-bias-pct P exits 1 if any decile or segment is biased by more than P%
Usage:
python residual_report.py preds.csv
python residual_report.py preds.csv --segment region --slice-by sqft year_built --top 5
python residual_report.py preds.csv --lower p10 --upper p90 --nominal 0.8
python residual_report.py preds.csv --max-abs-bias-pct 10 --json
Exit codes: 0 report printed (gates pass), 1 a bias gate failed, 2 usage or read error.
"""
import argparse
import csv
import json
import math
import sys
from collections import defaultdict
def num(v):
try:
x = float(v)
return x if math.isfinite(x) else None
except (TypeError, ValueError):
return None
def pearson(a, b):
n = len(a)
ma, mb = sum(a) / n, sum(b) / n
va = math.sqrt(sum((x - ma) ** 2 for x in a))
vb = math.sqrt(sum((y - mb) ** 2 for y in b))
return 0.0 if not va or not vb else sum((x - ma) * (y - mb) for x, y in zip(a, b)) / (va * vb)
def group_stats(pairs):
n = len(pairs)
err = [p - t for t, p in pairs]
st = sum(t for t, _ in pairs)
return {"n": n, "mae": sum(abs(e) for e in err) / n, "bias": sum(err) / n,
"bias_pct": (sum(err) / st * 100) if st else None,
"mean_true": st / n, "mean_pred": sum(p for _, p in pairs) / n}
def main():
ap = argparse.ArgumentParser(description="Residual diagnostics for regression predictions.", epilog=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("csv")
ap.add_argument("--true-col", default="y_true")
ap.add_argument("--pred-col", default="y_pred")
ap.add_argument("--segment")
ap.add_argument("--slice-by", nargs="*", default=[])
ap.add_argument("--lower")
ap.add_argument("--upper")
ap.add_argument("--nominal", type=float, default=0.8)
ap.add_argument("--top", type=int, default=5)
ap.add_argument("--max-abs-bias-pct", type=float)
ap.add_argument("--json", action="store_true")
a = ap.parse_args()
try:
with (sys.stdin if a.csv == "-" else open(a.csv, newline="", encoding="utf-8-sig")) as fh:
reader = csv.DictReader(fh)
cols = reader.fieldnames or []
need = [a.true_col, a.pred_col, a.segment, a.lower, a.upper, *a.slice_by]
miss = [c for c in need if c and c not in cols]
if miss:
raise ValueError(f"missing column(s) {miss}")
rows = [r for r in reader if num(r[a.true_col]) is not None and num(r[a.pred_col]) is not None]
except (OSError, ValueError) as e:
print(f"error: {e}", file=sys.stderr)
return 2
if len(rows) < 10:
print("error: fewer than 10 rows with numeric y_true and y_pred", file=sys.stderr)
return 2
y = [num(r[a.true_col]) for r in rows]
p = [num(r[a.pred_col]) for r in rows]
n = len(y)
err = [pi - yi for yi, pi in zip(y, p)]
ae = sorted(abs(e) for e in err)
mean_y = sum(y) / n
sst = sum((v - mean_y) ** 2 for v in y)
out = {"rows": n, "metrics": {
"mae": sum(ae) / n, "rmse": math.sqrt(sum(e * e for e in err) / n), "medae": ae[n // 2],
"bias": sum(err) / n, "r2": 1 - sum(e * e for e in err) / sst if sst else None,
"wape": sum(ae) / sum(abs(v) for v in y) if any(y) else None}}
order = sorted(range(n), key=lambda i: p[i])
dec = []
for d in range(10):
idx = order[d * n // 10:(d + 1) * n // 10]
if idx:
g = group_stats([(y[i], p[i]) for i in idx])
g["decile"] = d + 1
g["ratio_pred_true"] = g["mean_pred"] / g["mean_true"] if g["mean_true"] else None
dec.append(g)
out["deciles"] = dec
out["spread_corr"] = pearson(p, [abs(e) for e in err])
segs = {}
if a.segment:
groups = defaultdict(list)
for r, yi, pi in zip(rows, y, p):
groups[r[a.segment]].append((yi, pi))
segs[a.segment] = {k: group_stats(v) for k, v in sorted(groups.items())}
for c in a.slice_by:
vals = [(num(r[c]), yi, pi) for r, yi, pi in zip(rows, y, p) if num(r[c]) is not None]
vals.sort(key=lambda t: t[0])
m = len(vals)
segs[c] = {}
for qi in range(4):
part = vals[qi * m // 4:(qi + 1) * m // 4]
if part:
segs[c][f"Q{qi + 1} [{part[0][0]:g}, {part[-1][0]:g}]"] = group_stats([(t, q) for _, t, q in part])
out["segments"] = segs
if a.lower and a.upper:
lo = [num(r[a.lower]) for r in rows]
hi = [num(r[a.upper]) for r in rows]
inside = [l is not None and h is not None and l <= t <= h for l, h, t in zip(lo, hi, y)]
widths = [h - l for l, h in zip(lo, hi) if l is not None and h is not None]
out["intervals"] = {"coverage": sum(inside) / n, "nominal": a.nominal,
"mean_width": sum(widths) / len(widths) if widths else None}
worst = sorted(range(n), key=lambda i: -abs(err[i]))[:a.top]
out["worst"] = [{"row": i + 2, "y_true": y[i], "y_pred": p[i], "error": err[i]} for i in worst]
failed = []
if a.max_abs_bias_pct is not None:
for g in dec:
if g["bias_pct"] is not None and abs(g["bias_pct"]) > a.max_abs_bias_pct:
failed.append(f"decile {g['decile']} bias {g['bias_pct']:+.1f}%")
for col, gs in segs.items():
for k, g in gs.items():
if g["bias_pct"] is not None and abs(g["bias_pct"]) > a.max_abs_bias_pct:
failed.append(f"{col}={k} bias {g['bias_pct']:+.1f}%")
out["gate_failures"] = failed
if a.json:
print(json.dumps(out, indent=2, default=float))
return 1 if failed else 0
m = out["metrics"]
print(f"{n} rows MAE {m['mae']:.4g} RMSE {m['rmse']:.4g} MedAE {m['medae']:.4g} bias {m['bias']:+.4g} "
f"R2 {m['r2']:.4f} WAPE {m['wape']:.2%}")
print("\nCalibration by prediction decile (mean pred / mean true):")
print(" " + " ".join(f"D{g['decile']}:{g['ratio_pred_true']:.2f}" for g in dec if g["ratio_pred_true"]))
top_ratio = dec[-1]["ratio_pred_true"]
if top_ratio and top_ratio < 0.9:
print(f" the highest decile is under-predicted ({top_ratio:.2f}): typical of squared error on skewed targets, "
"shrinkage, or a missing feature for large cases")
if dec[0]["ratio_pred_true"] and dec[0]["ratio_pred_true"] > 1.1:
print(f" the lowest decile is over-predicted ({dec[0]['ratio_pred_true']:.2f}): regression to the mean")
sc = out["spread_corr"]
print(f"\n|residual| vs prediction correlation {sc:+.2f}" + (
": errors grow with the level (heteroscedastic) - log target / Gamma loss, report relative errors" if sc > 0.3 else ""))
for col, gs in segs.items():
print(f"\nBy {col}:")
for k, g in gs.items():
bp = "" if g["bias_pct"] is None else f" ({g['bias_pct']:+.1f}%)"
print(f" {k:<28} n={g['n']:<5} MAE {g['mae']:.4g} bias {g['bias']:+.4g}{bp}")
if "intervals" in out:
iv = out["intervals"]
print(f"\nInterval coverage {iv['coverage']:.1%} (nominal {iv['nominal']:.0%}), mean width {iv['mean_width']:.4g}"
+ (" <- too narrow" if iv["coverage"] < iv["nominal"] - 0.05 else
" <- wider than needed" if iv["coverage"] > iv["nominal"] + 0.05 else ""))
print(f"\nLargest errors (CSV line numbers): " + ", ".join(
f"line {w['row']}: true {w['y_true']:.4g} pred {w['y_pred']:.4g}" for w in out["worst"]))
for f in failed:
print(f"GATE FAILED: {f}")
return 1 if failed else 0
if __name__ == "__main__":
sys.exit(main())
anomaly-detection-model-builder
Build and evaluate anomaly detection for tabular, sensor and transaction data with few or no labels: define the anomaly type (point, contextual, collective), choose detectors (robust z-score/MAD and IQR rules, Mahalanobis with robust covariance, k-NN distance, LOF, Isolation Forest, one-class SVM, PCA reconstruction / autoencoders, forecasting residuals for time series), combine them by rank, explain each alert by its deviating features, set thresholds from an alert budget or the score distribution, and evaluate with precision@k, PR-AUC and recall on known incidents, or detector agreement when no labels exist. Includes pure-Python detectors (robust z, Mahalanobis, kNN, Isolation Forest, ensemble) and an evaluation script.
catboost-model-builder
Build CatBoost models that use its strengths correctly: raw categorical columns as cat_features (ordered target statistics, no manual encoding), text_features, Pool construction with NaN-safe categories, ordered boosting and has_time for temporal data, symmetric (oblivious) trees, overfitting detector with eval_set and use_best_model, losses for classification, regression (Quantile, Poisson, Tweedie) and ranking (YetiRank), class weights, monotone constraints, feature importance (PredictionValuesChange, LossFunctionChange) and SHAP values, GPU training and model export. Plans cat/text/numeric columns from a CSV and generates a runnable training script, and checks parameter sets for the conflicts CatBoost rejects (iterations vs n_estimators, Bayesian bootstrap with subsample, multiple class-weight options, depth > 16, use_best_model without eval_set).
categorical-encoding-selector
Choose and implement the right encoding for every categorical column given the model family and the data: one-hot (with infrequent-level grouping and handle_unknown), ordinal with an explicit order, cross-fitted target encoding (scikit-learn TargetEncoder, CatBoost ordered statistics, category_encoders inside CV), frequency/count encoding, feature hashing, native categorical support in LightGBM, HistGradientBoosting, XGBoost and CatBoost, prefix levels for codes (zip3, NAICS-2), and embeddings. Measures cardinality, rare-level share, unseen-level rate on a time-ordered holdout and target signal per column, emits the ColumnTransformer, and quantifies target-encoding leakage (in-sample vs out-of-fold vs holdout AUC).
classification-model-selector
Choose the right classification algorithm for tabular, text or mixed data and prove the choice with a fair bake-off: logistic regression (L2, L1, elastic net), linear SVM, naive Bayes, k-NN, RBF SVM, decision trees, random forest, extra trees, HistGradientBoosting, LightGBM, XGBoost, CatBoost, explainable boosting machines (EBM/GAM), points scorecards and MLPs, matched to data size, width, categorical and high-cardinality share, missingness, class imbalance, text, and constraints such as interpretability, monotonic constraints, calibrated probabilities, latency budgets and tiny deployable models. Ranks candidates with starting hyperparameters and cautions, then generates a bake-off script running every candidate on identical CV folds with per-fold metrics, fit time and prediction latency.
clustering-algorithm-selector
Choose the right clustering algorithm and check that clustering makes sense at all: cluster tendency (Hopkins statistic with a correlated-features caveat, nearest-neighbour isolation), then k-means / MiniBatchKMeans, Gaussian mixtures with BIC, agglomerative (Ward, average, complete, single) with dendrograms, DBSCAN, HDBSCAN, OPTICS, spectral clustering, BIRCH, mean shift, k-prototypes and k-modes for mixed or categorical data, and Gower distance, matched to data size, expected shapes, noise, varying density, known k, soft memberships, hierarchy and streaming. Covers preprocessing (scaling, encoding, dimensionality, distance metrics), validation (internal indices, bootstrap stability, external labels, business profiling) and assigning new points.
collaborative-filtering-implementer
Implement collaborative filtering recommenders that beat popularity: item-based kNN with cosine, Jaccard or asymmetric similarity, shrinkage, neighbourhood size, positive-only or mean-centred rating signals and 'because you liked X' explanations; implicit ALS with confidence weighting, BPR pairwise ranking and the closed-form EASE model in plain NumPy, tuned on a temporal validation split by recall@k and NDCG@k and refit on all data; plus user-kNN, explicit matrix factorisation, the implicit/LightFM libraries, hyperparameter ranges, cold-start fallbacks, scaling with sparse matrices and approximate nearest neighbours. Produces top-k lists in the format of the recommendation evaluator.